Pith. sign in

REVIEW 4 major objections 4 minor 43 references

Constrained Pareto Set Identification with Bandit Feedback

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper introduces constrained Pareto Set Identification with explainability, a new bandit pure-exploration problem where the learner must find the Pareto-optimal arms among those satisfying known linear feasibility constraints and…

desk verdict A solid new problem and algorithm for constrained Pareto set identification, but the 'near-optimal' claim is proven only under a norm-equivalence condition and the upper bound has an unmatched additive term. read the letter →

arxiv 2506.08127 v1 pith:JZ4EBQQQ submitted 2025-06-09 stat.ML cs.LG

classification stat.MLcs.LG
keywords constrainedParetosetidentificationmulti-objectivebanditspureexplorationfixedconfidencesamplecomplexityfeasibilityconstraintsexplainability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces a new bandit pure-exploration problem: given K arms with unknown mean vectors in R^d and a known polyhedron of acceptable means, find the Pareto-optimal arms among those whose means lie in the polyhedron, and for every rejected arm state whether it is infeasible or dominated by a feasible arm. The central claim is that this explainable version can be solved near-optimally: the proposed algorithm e-cAPE is delta-correct, and its sample complexity is O(C*_M(nu) log(K C*_M(nu)/delta)) plus an additive feasibility term, matching an information-theoretic lower bound of $\Omega$(C*_M(nu) log(1/delta)) up to constants on a constructed family of Gaussian instances. The complexity measure C*_M(nu) is a minimum over valid explanations of the cost of certifying feasibility and Pareto dominance for each arm, and it never exceeds the cost of the natural two-stage approach that first identifies feasible arms and then their Pareto set. The paper also proposes an asymptotically optimal algorithm for the non-explainable version and reports empirical gains on clinical-trial data.

What carries the argument

The central object is e-cAPE, a top-two confidence-bound algorithm. It maintains, for every pair of arms (i,j), upper and lower confidence bounds on M(i,j) = max_c [mu_i^c - mu_j^c] and m(i,j) = min_c [mu_j^c - mu_i^c], which capture Pareto dominance, and confidence bounds on eta_i, the distance of arm i's mean to the polyhedron P, which captures feasibility. The stopping rule monitors two statistics: Z1(t), which certifies that the empirical feasible Pareto set O_t is truly feasible and that its arms are not dominated by each other, and Z2(t), which certifies that every arm outside O_t is either infeasible or dominated by an arm in F_t union G_t. The sampling rule pulls a leader b_t that minimizes the current stopping statistic and a challenger c_t that is closest to dominating b_t, thereby focusing samples on the arms that most delay a valid explanation. The complexity C*_M(nu) is the minimum over all valid partitions (S,I) of the sum of inverse squared gaps that appear in the analysis of the stopping time.

What would settle it

Build an instance where condition (53) fails, for example one suboptimal arm dominated by two Pareto-optimal arms with incomparable margins in dimension 2, so that the Euclidean perturbation norm ||s_i||_2 exceeds the sup-norm gap Delta_i by a factor near sqrt(2); then run any delta-correct algorithm, including e-cAPE, as delta tends to 0 and measure the ratio of expected stopping time to C*_M(nu) log(1/delta). The paper's unrestricted near-optimality claim predicts a limiting ratio of at least 1/8, while a ratio below that would show the lower bound does not hold for general instances.

Watch

Extended reading notes

Core claim

The core discovery is that explainable constrained Pareto set identification has a well-defined, instance-dependent sample complexity given by C*_M(nu) = min_{(S,I) in M} C(nu,S,I), where C sums inverse squared feasibility distances 1/$eta_i^{2}$ and Pareto gaps 1/$Delta_i^{2}$ over arms that must be certified feasible, infeasible, or dominated; e-cAPE achieves this complexity up to a logarithmic factor. The paper proves that e-cAPE is delta-correct (Theorem 4.4) with the stated sample-complexity bound, and that on a restricted family of Gaussian instances any delta-correct algorithm needs at least C*_M(nu) log(1/delta)/8 samples in the limit delta->0 (Theorem 4.5), so e-cAPE is near-optimal on those instances. The efficiency comes from balancing feasibility detection and Pareto identification through confidence bounds, and from the freedom to choose the easiest valid explanation for each rejected arm.

Load-bearing premise

The near-optimality proof assumes that the Euclidean perturbation cost used in the lower-bound construction equals the sup-norm Pareto gaps used in the upper-bound complexity, an equality that holds exactly only on the restricted instance family of Appendix E.4, condition (53).

Editorial extensions

If this is right

  • The leading sample-complexity term of e-cAPE, C*_M(nu), is never larger than the two-stage baseline's H_PSI(nu) + H_F(nu), so the explainable formulation is not more expensive than first certifying feasibility and then Pareto optimality.
  • When feasibility distances eta_i are large, e-cAPE's guarantee reduces to the unconstrained Pareto Set Identification bound of Auer et al. (2016) and Kone et al. (2023); when Pareto gaps are large, it reduces to the feasibility-detection bound of Katz-Samuels & Scott (2018).
  • With law-of-the-iterated-logarithm confidence bounds, e-cAPE satisfies a high-probability bound of the form O(C_M(nu) log(Kd log(C_M(nu))/delta)).
  • On the Gaussian families constructed in Appendix E.4 under condition (53), any delta-correct algorithm needs at least C*_M(nu)/8 * log(1/delta) samples as delta->0, so e-cAPE is near-optimal there.
  • The racing algorithm for e-cPSI is strictly suboptimal in the worst case: when all arms are feasible and only one is Pareto-optimal, racing pays an extra 1/eta_1^2 feasibility cost that e-cAPE avoids.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The algorithm's efficiency engine is the minimum over valid explanations: an arm near the feasibility boundary that is also dominated can be certified as dominated at the cheaper of the two costs, which is likely why e-cAPE wins on the clinical-trial instances where the 75mg arm is both boundary-near and suboptimal.
  • The near-optimality claim is only proven for instances satisfying condition (53); a natural, testable extension is a matching lower bound for general instances, which would probably require a complexity measure using sup-norm Pareto gaps rather than Euclidean perturbation costs.
  • The same confidence-bound machinery could handle time-varying or random polyhedra—for instance, when feasibility constraints are estimated from the data themselves—by recomputing eta_i(t) and the Pareto gaps online at each round.
  • The empirical result that e-cAPE sometimes beats the asymptotically optimal Game-cPSI even for the non-explainable task suggests that the moderate-confidence regime, not the delta->0 limit, is where the explainable formulation's flexibility pays off; this could be tested systematically across a broader benchmark.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces constrained Pareto Set Identification (cPSI) and an explainable variant (e-cPSI) in a fixed-confidence multivariate bandit setting, where the learner must identify the Pareto set of arms whose means lie in a known polyhedron. The main algorithmic contribution is e-cAPE, a confidence-bound-based algorithm that combines feasibility detection with Pareto-set identification. Theorem 4.4 states that e-cAPE is delta-correct with an expected sample complexity of O(C*_M(nu) log(K C*_M(nu)/delta)) plus an additive feasibility term H(nu,F^c). Theorem 4.5 / Theorem E.1 gives a lower bound of Omega(C*_M(nu) log(1/delta)) on a restricted family of Gaussian instances, on which the algorithm is claimed to be near-optimal. The paper also sketches an asymptotically optimal game-based algorithm for cPSI and reports empirical comparisons on clinical-trial-inspired datasets.

Significance. If the results hold as stated, the paper extends pure-exploration Pareto-set identification to feasibility-constrained problems and introduces a useful 'explainability' requirement in which every rejected arm must be classified as infeasible or dominated. The correctness proof of the stopping rule (Lemma 4.3 and Appendix C.1) is careful, and the algorithm is concrete and reproducible. The lower-bound construction via change of measure is a standard and appropriate technique, and the empirical evaluation includes meaningful baselines (two-stage, racing, uniform). The strengths are the explicit finite-confidence algorithm, a nontrivial complexity measure C*_M(nu), and the attempt to match it with a lower bound. However, the advertised 'near-optimal' claim is currently stronger than what the theorems establish: the lower bound is proved only on a specially constructed family, and the upper-bound proof has a constant-factor inconsistency. The central algorithmic idea is defensible, but the paper's headline claim needs qualification and the lower-bound proof needs to be made self-contained.

major comments (4)
  1. [Theorem 4.4 and Appendix C.2.2] The stated upper bound is inconsistent with its proof. In the proof, after Lemma D.1 the authors define \tilde C(\nu,S,I) = 64 C(\nu,S,I) \alpha and obtain T^*(S,I) = 2\tilde C \log(\tilde C(4k_1Kd/\delta)^{1/\alpha}), which gives a prefactor 128 \alpha C(\nu,S,I) and a log argument 64 \alpha C(\nu,S,I). However, Theorem 4.4 states 128 C^*_M(\nu) \log(64 C^*_M(\nu)(4k_1Kd/\delta)^{1/\alpha}) with no \alpha in either place. In addition, the proof line 'a=64 \sigma^2 C(\nu,S,I)\alpha' appears to contain a spurious \sigma^2, since \sigma^2 is already absorbed in C. The theorem should be corrected to match the derivation.
  2. [Abstract, Section 1.2, Theorem 4.5, Appendix E.4] The abstract and the conclusion state that the paper proves near-optimality of the sample complexity of e-cAPE, but the formal lower bound is only proved on a restricted family. Theorem 4.5 states 'There exists a class eD such that ...', and Appendix E.4 constructs eD. Moreover, the constructed family has no feasible suboptimal arms (for these instances every feasible arm is Pareto-optimal, so SubOpt is empty), and for Pareto-optimal arms the Euclidean quantity e\delta_i^+ equals 2\alpha while the sup-norm gap \delta_i^+ equals \alpha, so even condition (53) is not literally satisfied. The lower bound therefore exercises only the O^* and I terms of C^*_M(\nu), not the S terms. The unqualified 'near-optimal' claim in the abstract, contributions, and conclusion should be replaced by a statement that e-cAPE is near-optimal on the constructed classes, and the gap with the general lower bound of Proposition 3.4 should be acknowledged.
  3. [Appendix E.3, after Eq. (52)] The lower-bound proof contains an unproved step for the gap \delta_i^-(S) of an optimal arm. The text says 'As justified in Remark 18 of Auer et al. (2016), \delta_i^-(S), can be compensated by both \delta_i^+(S) and \Delta_j^* for some arms j\in SubOpt (in our case). So we focus on matching the gaps in \delta_i^+'. This is a load-bearing claim because C(\nu,S,I) uses \Delta_i(O^*\cup S)=\min(\delta_i^+, \delta_i^-(S)). The lower bound derived in Section E.2 only gives a bound involving \min(e\delta_i^+, \eta_i), not \min(\Delta_i(O^*\cup S), \eta_i). To make Theorem E.1 self-contained, the authors must either prove the compensation argument explicitly or restrict the statement of the lower bound to instances in which \delta_i^-(S) is vacuous or otherwise dominated by the terms that are matched.
  4. [Theorem 4.4, displayed bound and definition of H(nu,F^c)] The additive term 4H(\nu,F^c) is not matched by the lower bound, and its definition is notationally confusing. The theorem defines H(\nu,F^c) as a sum over F^c \cup O^*, not over F^c. More importantly, the proof obtains E[\tau] \le T^*(S,I) + \Lambda_\alpha with T^*(S,I)=2\tilde C\log(\tilde C b)+4H(\nu,I), and then uses H(\nu,I) \le H(\nu,F^c). Since the lower bound has no analogous additive term and the lower-bound family has no feasible suboptimal arms, finite-confidence near-optimality in the regime of fixed moderate \delta is not established; the paper should state precisely that the matching is asymptotic in \delta \to 0 or require a lower bound that includes the feasibility term.
minor comments (4)
  1. [Theorem E.1 statement] The statement reads 'E_\nu [\tau] min_{(S',I') \in M} C(\nu,S',I') g(\delta)/4' with no inequality; it should be 'E_\nu[\tau] \ge \min_{(S',I')} C(\nu,S',I') g(\delta)/4'.
  2. [Theorem 4.4, definition of H(\nu,F^c)] The expression '\log(5d/d)' appears to be a typo: the proof uses quantities such as f(T,5^{-d}d) and f(T,\delta\cdot d/5^d), so the correct term should presumably be \log(5^d/d), not \log(5d/d).
  3. [Appendix E.4] The construction says 'd \ge 1 (arbitrary)' but then uses coordinates 1 and 2; the statement should say d \ge 2.
  4. [Section 4.1, Algorithm 1] In Algorithm 1, the description of the leader b_t as the minimizer of (Z^F_1(t) if Z^{PS}_1(t) \ge Z^F_1(t), and Z^{PS}_1(t) otherwise) is terse; a parenthetical restating that this is min(Z^F_1(t), Z^{PS}_1(t)) would improve readability.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: C*_M appears in both bounds but the lower bound is an independent change-of-measure argument; the near-optimality claim is qualified by condition (53), not circularly derived.

full rationale

The core derivation is not circular. The upper bound (Theorem 4.4, Appendix C.2) directly analyzes e-cAPE: on the good event, Proposition C.3 shows each non-stopped round samples an under-explored arm in W_t(S,I), and the sum of inverse-squared gaps yields O(C*_M(ν) log(KC*_M(ν)/δ)) + 4H(ν,F^c) + Λ_α. The lower bound (Theorem 4.5, Appendix E) uses the change-of-distribution Lemma E.2 and the δ-correctness definition only; it never assumes e-cAPE's sampling rule. It lower-bounds E[N_{τ,i}] by 2/η_i^2, 2/||s_i||_2^2, or 2/min(η_i, eδ+_i)^2 on alternative Gaussian instances, then maps these to C(ν,S,I). Thus the appearance of C*_M on both sides is a genuine matching on the constructed family, not a fitted parameter renamed as a prediction. The only self-citation, to the authors' prior APE work (Kone et al. 2023), supplies the confidence-region design template and standard PSI gap definitions; Lemma 4.3 and Theorem 4.4 are proved from scratch in Appendices C.1-C.2, so this citation is not load-bearing. The skeptic's concern is a qualification gap rather than circularity: Appendix E.3 admits that outside condition (53) the Euclidean quantities in (46) and (50) can exceed the sup-norm gaps by up to √d or √|Ω_i|, and Theorem 4.4's additive feasibility term 4H(ν,F^c) is not matched by the lower bound. Theorem 4.5 is therefore stated only for the restricted family eD, so the abstract's unqualified 'near-optimal' overreaches. The proof also depends on Auer et al.'s Remark 18 to compensate δ^-_i(S) gaps without a self-contained argument. These are missing or overstated support, not self-referential reductions.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central results rest on standard bandit concentration and change-of-measure machinery, plus the problem-domain assumption of a known linear constraint polyhedron. No ad hoc entities are introduced. The only hand-chosen quantities are the universal constants k1 and α in the confidence bounds, which affect constants but not the problem-dependent complexity order.

free parameters (2)
  • k1 = k1 > 1 + 1/(α-1)
    A universal constant in the confidence bound f(t,δ) = log(4 k1 K d t^α/δ). Chosen by hand to make the union bound over time in Lemma D.3 hold; not fitted to data.
  • α = α > 2
    Exponent in the confidence bound log(t^α/δ); chosen so that the series Σ t^{-α} converges, implying the error probability sum is O(δ). Not fitted to data.
assumptions (5)
  • domain assumption Arms are independent with i.i.d. samples; each arm is marginally σ-subGaussian and norm-σ_u-subGaussian (or multivariate Gaussian with identity covariance for lower bounds).
    Stated in Section 1.1 and used in Lemma 4.2, Lemma D.3, and the lower-bound proofs in Section 3 and Appendix E.
  • domain assumption The feasibility polyhedron P = {x : Ax ≤ b} is known to the learner.
    This is part of the problem definition; distances to P and projections onto P are used in the algorithm, its analysis (Lemma 4.2, Lemma D.2), and lower-bound constructions.
  • standard math Time-uniform confidence inequalities for subGaussian and norm-subGaussian vector means (Garivier 2013; Kaufmann & Koolen 2021; ε-net covering arguments).
    Used in Lemma D.3 to control the event E that the confidence bounds hold simultaneously over time and arms.
  • standard math Change-of-distribution lemma for bandits (Kaufmann et al. 2016, Lemma E.2).
    Used in Appendix E to transform the algorithm's correctness under alternative instances into lower bounds on expected pulls per arm.
  • standard math Convex quadratic programs can be solved in polynomial time (Ye & Tse 1989), and the valid-map enumeration of Crepon et al. (2024) bounds the number of subproblems in the best-response computation.
    Used in Appendix B.3 (Lemma B.5) to establish that Game-cPSI's per-round computational cost is polynomial in K.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Constrained Pareto Set Identification with Bandit Feedback." pith.science (2026). https://pith.science/paper/JZ4EBQQQ

@misc{pith2026250608127,
  author       = {Pith},
  title        = {Pith review of: Constrained Pareto Set Identification with Bandit Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZ4EBQQQ}},
  note         = {Machine review of arXiv:2506.08127}
}
abstract

In this paper, we address the problem of identifying the Pareto Set under feasibility constraints in a multivariate bandit setting. Specifically, given a $K$-armed bandit with unknown means $\mu_1, \dots, \mu_K \in \mathbb{R}^d$, the goal is to identify the set of arms whose mean is not uniformly worse than that of another arm (i.e., not smaller for all objectives), while satisfying some known set of linear constraints, expressing, for example, some minimal performance on each objective. Our focus lies in fixed-confidence identification, for which we introduce an algorithm that significantly outperforms racing-like algorithms and the intuitive two-stage approach that first identifies feasible arms and then their Pareto Set. We further prove an information-theoretic lower bound on the sample complexity of any algorithm for constrained Pareto Set identification, showing that the sample complexity of our approach is near-optimal. Our theoretical results are supported by an extensive empirical evaluation on a series of benchmarks.

Figures

Figures reproduced from arXiv: 2506.08127 by the authors.

Figure 1
Figure 1. Example of an instance of constrained PSI. The com￾plexity of a two-stage approach will scale with 1/η2 1 + 2/ε2 P + 2 i=1 1/∆i({1, 2}) 2 while e-cAPE’s complexity scales with 1/η2 1 + P3 i=1 1/∆i({1, 2, 3}) 2 when ε ≪ 1. Moreover, we remark that when the arms are very far from the boundaries of P, i.e., ηi ≫ 1, e-cAPE matches the bounds of Auer et al. (2016) and Kone et al. (2023) for PSI on F. Similarly, when the … view at source ↗
Figure 2
Figure 2. Average response for each dosage level. The feasible region is defined by an effectiveness of at least 40% and non-toxicity IgG 7 8 9 10 11 NT50 5 6 7 8 cell. resp o nse 3.5 4.0 4.5 5.0 5.5 Feasible Region O * S [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 4
Figure 4. Empirical sample complexity averaged on 500 runs for the secukinumab trial (left) and the covboost trial (right). Through our simulations from real-world and synthetic data, we demonstrate the superior empirical performance of e-cAPE, highlighting its efficiency for constrained Pareto Set Identification. While our work ensures constraint satisfaction in the fi￾nal recommendation, an exciting avenue for future resear… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Constrained PSI instances with (left to right): ordered polyhedron, cube, simplex. The green area denotes the feasible region. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_5.png]
Figure 6
Figure 6. Figure 6: Constrained PSI instance (left) and distribution of the empirical sample complexity (right) averaged over 250 runs. Sample complexity Quite surprisingly, on that instance the sample complexity of e-cPSI is actually (slightly) smaller than of cPSI, despite the fact that…
Figure 7
Figure 7. Figure 7: Constrained 5-armed PSI instance on the left and em￾pirical distribution of the sample complexity (right). 0.4 0.5 0.6 0.7 fx 0.30 0.35 0.40 0.45 0.50 0.55 0.60 fy O∗ S e-cAPE A-A R-CP U 10,000 20,000 30,000 40,000 [PITH_FULL_IMAGE:figures/full_fig_p037_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 42 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Constrained multi-objective optimization algorithms: Review and comparison with application in reinforced concrete structures

    Afshari, H., Hare, W., and Tesfamariam, S. Constrained multi-objective optimization algorithms: Review and comparison with application in reinforced concrete structures. Applied Soft Computing, 2019

  3. [3]

    A learning-based approach to the automated design of mpsoc networks

    Almer, O., Topham, N., and Franke, B. A learning-based approach to the automated design of mpsoc networks. In Proceedings of the 24th International Conference on Architecture of Computing Systems. Springer-Verlag, 2011

  4. [4]

    Linear stochastic bandits under safety constraints

    Amani, S., Alizadeh, M., and Thrampoulidis, C. Linear stochastic bandits under safety constraints. In Advances in Neural Information Processing Systems, 2019

  5. [5]

    and Bubeck, S

    Audibert, J.-Y. and Bubeck, S. Best arm identification in multi-armed bandits. In COLT - 23th Conference on Learning Theory , 2010

  6. [6]

    Pareto front identification from stochastic bandit feedback

    Auer, P., Chiang, C.-K., Ortner, R., and Drugan, M.-M. Pareto front identification from stochastic bandit feedback. In Proceedings of the 19th International Conference on Artificial Intelligence and Statistics. PMLR , 2016

  7. [7]

    Sequential learning of the P areto front for multi-objective bandits

    C repon, E., Garivier, A., and M Koolen, W. Sequential learning of the P areto front for multi-objective bandits. In Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research. PMLR, 2024

  8. [8]

    D., and Koolen, W

    De Rooij, S., Van Erven, T., Gr\" u nwald, P. D., and Koolen, W. M. Follow the leader if you can, hedge if you must. J. Mach. Learn. Res., 2014

Show all 43 references
  1. [9]

    and Koolen, W

    Degenne, R. and Koolen, W. Pure exploration with multiple correct answers. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019

  2. [10]

    Gamification of pure exploration for linear bandits

    Degenne, R., M \' e nard, P., Shang, X., and Valko, M. Gamification of pure exploration for linear bandits. In International Conference on Machine Learning (ICML) , 2020

  3. [11]

    and Nowe, A

    Drugan, M.-M. and Nowe, A. Designing multi-objective multi-armed bandits algorithms: A study. In The 2013 International Joint Conference on Neural Networks (IJCNN), 2013

  4. [12]

    Action elimination and stopping conditions for the multi-armed bandit and reinforcement learning problems

    Even-Dar, E., Mannor, S., and Mansour, Y. Action elimination and stopping conditions for the multi-armed bandit and reinforcement learning problems. Journal of Machine Learning Research, 2006

  5. [13]

    Informational confidence bounds for self-normalized averages and applications

    Garivier, A. Informational confidence bounds for self-normalized averages and applications. In 2013 IEEE Information Theory Workshop (ITW). IEEE, September 2013

  6. [14]

    and Kaufmann, E

    Garivier, A. and Kaufmann, E. Optimal best arm identification with fixed confidence. In 29th Annual Conference on Learning Theory. PMLR, 2016

  7. [15]

    Exploring multi-objective exercise recommendations in online education systems

    Huang, Z., Liu, Q., Zhai, C., Yin, Y., Chen, E., Gao, W., and Hu, G. Exploring multi-objective exercise recommendations in online education systems. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management. Association for Computing Machi...

  8. [16]

    and Nowak, R

    Jamieson, K. and Nowak, R. Best-arm identification algorithms for multi-armed bandits in the fixed confidence setting. In 2014 48th Annual Conference on Information Sciences and Systems ( CISS ) , 2014

  9. [17]

    lil' ucb : An optimal exploration algorithm for multi-armed bandits

    Jamieson, K., Malloy, M., Nowak, R., and Bubeck, S. lil' ucb : An optimal exploration algorithm for multi-armed bandits. In Proceedings of The 27th Conference on Learning Theory. PMLR, 2014

  10. [18]

    and Turnbull, B

    Jennison, C. and Turnbull, B. W. Group sequential tests for bivariate response: interim analyses of clinical trials with both efficacy and safety endpoints. Biometrics, 1993

  11. [19]

    PAC subset selection in stochastic multi-armed bandits

    Kalyanakrishnan, S., Tewari, A., Auer, P., and Stone, P. PAC subset selection in stochastic multi-armed bandits. In Proceedings of the 29th International Coference on International Conference on Machine Learning. Omnipress, 2012

  12. [20]

    M., Y ld r m, Y

    Karag \"o zl \"u , E. M., Y ld r m, Y. C., Ararat, C., and Tekin, C. Learning the P areto set under incomplete preferences: Pure exploration in vector bandits. In Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, Proceedings of Machine...

  13. [21]

    and Scott, C

    Katz-Samuels, J. and Scott, C. Feasible arm identification. In Proceedings of the 35th International Conference on Machine Learning. PMLR , 2018

  14. [22]

    and Scott, C

    Katz-Samuels, J. and Scott, C. Top feasible arm identification. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics. PMLR , 2019

  15. [23]

    Analysis of bayesian and frequentist strategies for sequential resource allocation

    Kaufmann, E. Analysis of bayesian and frequentist strategies for sequential resource allocation . Thesis, T \'e l \'e com ParisTech , 2014

  16. [24]

    and Koolen, W.-M

    Kaufmann, E. and Koolen, W.-M. Mixture martingales revisited with applications to sequential tests and confidence intervals. Journal of Machine Learning Research, 2021

  17. [25]

    On the complexity of best-arm identification in multi-armed bandit models

    Kaufmann, E., Capp \'e , O., and Garivier, A. On the complexity of best-arm identification in multi-armed bandit models. Journal of Machine Learning Research, 2016

  18. [26]

    Adaptive algorithms for relaxed pareto set identification

    Kone, C., Kaufmann, E., and Richert, L. Adaptive algorithms for relaxed pareto set identification. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

  19. [27]

    Pareto set identification with posterior sampling, 2024 a

    Kone, C., Jourdan, M., and Kaufmann, E. Pareto set identification with posterior sampling, 2024 a

  20. [28]

    Bandit pareto set identification in a multi-output linear model

    Kone, C., Kaufmann, E., and Richert, L. Bandit pareto set identification in a multi-output linear model. In Seventeenth European Workshop on Reinforcement Learning, 2024 b

  21. [29]

    Z., Luo, Q., Mallipeddi, R., Suganthan, P

    Kumar, A., Wu, G., Ali, M. Z., Luo, Q., Mallipeddi, R., Suganthan, P. N., and Das, S. A benchmark-suite of real-world constrained multi-objective optimization problems and some baseline results. Swarm and Evolutionary Computation, 2021

  22. [30]

    and Szepesvári, C

    Lattimore, T. and Szepesvári, C. Bandit Algorithms. Cambridge University Press, 2020

  23. [31]

    Constrained multi-objective bayesian optimization through optimistic constraints estimation, 2024

    Li, D., Zhang, F., Liu, C., and Chen, Y. Constrained multi-objective bayesian optimization through optimistic constraints estimation, 2024

  24. [32]

    A pareto-efficient algorithm for multiple objective optimization in e-commerce recommendation

    Lin, X., Chen, H., Pei, C., Sun, F., Xiao, X., Sun, H., Zhang, Y., Ou, W., and Jiang, P. A pareto-efficient algorithm for multiple objective optimization in e-commerce recommendation. In Proceedings of the 13th ACM Conference on Recommender Systems, RecSys '19. Association for...

  25. [33]

    Efficacy and safety of secukinumab in patients with rheumatoid arthritis: a phase ii, dose-finding, double-blind, randomised, placebo controlled study

    Mark C, G., Patrick, D., Hanno B, R., Jerzy, S., Eva, D., Vadim, M., Jacob A, A., Sang-Heon, L., Christine E, C., Herbert, K., Takashi, I., Sophie, H., and Shephard, M. Efficacy and safety of secukinumab in patients with rheumatoid arthritis: a phase ii, dose-finding, double-b...

  26. [34]

    Bandit based optimization of multiple objectives on a music streaming platform

    Mehrotra, R., Xue, N., and Lalmas, M. Bandit based optimization of multiple objectives on a music streaming platform. KDD '20. Association for Computing Machinery, 2020

  27. [35]

    Gradient ascent for active exploration in bandit problems, 2019

    M\'enard, P. Gradient ascent for active exploration in bandit problems, 2019

  28. [36]

    Munro, A.-P.-S., Janani, L., Cornelius, V., and et al. Safety and immunogenicity of seven COVID -19 vaccines as a third dose (booster) following two doses of ChAdOx 1 nCov -19 or BNT 162b2 in the UK ( COV - BOOST ): a blinded, multicentre, randomised, controlled, phase 2 trial...

  29. [37]

    Dealing with misspecification in fixed-confidence linear top-m identification

    R\' e da, C., Tirinzoni, A., and Degenne, R. Dealing with misspecification in fixed-confidence linear top-m identification. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2021

  30. [38]

    and Vershynin, R

    Rudelson, M. and Vershynin, R. Hanson-wright inequality and sub-gaussian concentration, 2013

  31. [39]

    Fast pure exploration via frank-wolfe

    Wang, P.-A., Tzeng, R.-C., and Proutiere, A. Fast pure exploration via frank-wolfe. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2021

  32. [40]

    and Tse, E

    Ye, Y. and Tse, E. An extension of karmarkar projective algorithm for convex quadratic programming. Math. Program., July 1989

  33. [41]

    Information-directed selection for top-two algorithms

    You, W., Qin, C., Wang, Z., and Yang, S. Information-directed selection for top-two algorithms. In Proceedings of Thirty Sixth Conference on Learning Theory, Proceedings of Machine Learning Research. PMLR, 2023

  34. [42]

    Active learning for multi-objective optimization

    Zuluaga, M., Sergent, G., Krause, A., and Püschel, M. Active learning for multi-objective optimization. In Proceedings of the 30th International Conference on Machine Learning. PMLR , 2013

  35. [43]

    e-pal: An active learning approach to the multi-objective optimization problem

    Zuluaga, M., Krause, A., and P \"u schel, M. e-pal: An active learning approach to the multi-objective optimization problem. Journal of Machine Learning Research, 2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.